Continual learning, adaptive skills, and safe runtime improvement

A reported Amazon weight-memory study says on-policy self-distillation can improve user-fact recall over retrieval, with additional gains from text-store verification; its eight-user evaluation and reported forgetting limit generalization. P²-MOPD claims faster recovery in a distillation setup using expert task-vector priors, but benchmark context and independent evidence are missing. A self-improving-agent resource hub provides a useful taxonomy but no new evidence, while Inkling 2 Small remains an unsubstantiated teaser.